Loader

Loader is Chibi-Nodes' all-in-one starting point: it merges what would normally be three separate ComfyUI nodes - CheckpointLoaderSimple, CLIPSetLastLayer, and EmptyLatentImage - into one. Instead of placing three nodes and wiring MODEL/CLIP/VAE across each one individually, you get all four typed outputs (MODEL, VAE, CLIP, LATENT) out of a single box. For a first workflow, that's a genuinely meaningful reduction in things you can mis-wire.

Pair it with this pack's own Prompts node (which takes a CLIP input and gives you CONDITIONING back) and SimpleSampler (which takes MODEL, CONDITIONING × 2, and LATENT), and you've got a complete, minimal txt2img graph in three Chibi nodes total: Loader → Prompts → SimpleSampler.

The inputs and outputs that matter

• Checkpoint - the dropdown of installed checkpoints, same list as ComfyUI's stock loader draws from models/checkpoints.
• Vae - "Included" uses the checkpoint's own baked-in VAE; anything else installed in models/vae also appears here if you want to override it. Some checkpoints - certain older SD1.5 merges especially - ship without a well-tuned baked VAE and look washed out or grey until you swap in a proper external one, so this dropdown existing at all is worth knowing about even if you leave it on "Included" most of the time.
• stop_at_clip_layer - this is ComfyUI's CLIP-skip control (the same thing as CLIPSetLastLayer), ranging -24 to -1 with -1 as the default (off). If that number means nothing to you yet: it originates from NovelAI's 2022 choice to condition on CLIP's second-to-last layer instead of its last, and every anime-lineage checkpoint since - Illustrious, Pony, NoobAI - inherited the convention of running at -2. Set it to -2 if your checkpoint's model card recommends "CLIP skip 2" (ComfyUI counts backward from -1, where A1111-style UIs count forward from 1, so A1111's "2" is ComfyUI's -2). Leave it at -1 for realistic SD 1.5 checkpoints and don't bother touching it at all on SDXL - SDXL already conditions on the penultimate layer internally, so the setting is mostly a no-op there, and it doesn't exist as a concept at all on 2026-era models using Qwen3 or Mistral text encoders instead of CLIP.
• width / height - default 512, stepping by 8 (not 1) - that step size isn't arbitrary, it's the standard SD-family constraint that latent dimensions need to be multiples of 8.
• batch_size - 1 to 4096. The ceiling is generous; you won't realistically run anywhere near it unless you've got serious VRAM and a specific reason to.

Four outputs - MODEL, VAE, CLIP, LATENT - cover everything a basic txt2img sampler chain needs.

How to install it

• ComfyUI Manager: search "ComfyUI-Chibi-Nodes", install, restart.
• Manual: cd ComfyUI/custom_nodes && git clone https://github.com/chibiace/ComfyUI-Chibi-Nodes/, restart.

No extra downloads beyond checkpoints/VAEs you'd need regardless of which loader node you use. It appears under Chibi-Nodes in the menu. The pack overall is a small personal utility collection - the README calls these "experimental nodes" and admits version 2 shipped "with more bugs" - worth keeping as general calibration, though this particular node is just recombining ComfyUI's own well-tested loading logic.

Common issues

If the Checkpoint dropdown is empty, you don't have anything in models/checkpoints yet - that's an installation issue, not a node bug.

The stop_at_clip_layer field is the one setting here people genuinely misuse: it's a real, load-bearing setting on SD 1.5 anime checkpoints and on the Illustrious/Pony/NoobAI family (where a mismatch between what the model expects and what you set is a standard entry on the "why doesn't my LoRA look right" checklist), but it's cargo-culted onto models where it does nothing - don't chase quality by cranking it to -3 or -4 on an SDXL checkpoint expecting an improvement; each step just strips more specificity out of the conditioning, and past -2 you're usually just making things worse for no benefit.